Inception Network Video Under Case Studies (Week 2)

Need clarification on Inception Network video.

Does Dr. Ng arbitrarily chose the number of channels for 1x1 convolutions? For example, he chose 16 channels for 1x1 convolutions on the video. Why is that?


I think its probably the same in the original inception model and if not then its similar to the number of channels used in similar application, its good to always choose a power of 2, for computational performance.